Implementation And Research Of Edge Computing For Deep Learning

In recent years,cloud computing has played an essential role on the internet.From personal consumer products to industrial products,there are cloud computing services as the support behind the products.Cloud computing has caused earth-shaking changes in various industries.The complexity and diversity of terminal devices are also increasing.Using traditional cloud computing to achieve low latency and low power consumption is a considerable challenge for terminal devices.Edge computing can make up for the deficiencies of cloud computing in these aspects.Deep learning has found wide usage in recent years.Deep learning technologies are indispensable from speech recognition to image recognition,and they are computationally intensive,complex and very tough to deploy.Deploying deep learning algorithms is more challenging in terminal devices,edge computing can help in this regards.This thesis deeply study the methods of deep learning training and computation in edge computing environment,analyzed the significance and challenges of deploying deep learning in the edge computing environment.This thesis proposes a computation offloading approach based on the RSSI strength of the terminal devices,and aims to reduce the power consumption of terminal devices.This thesis builds a hardware platform to verify the effectiveness of this computation offloading approach.This hardware platform uses a raspberry pi equipped with a4 G network card as a terminal device,and it uses an Alibaba cloud server as an edge server.In order to test the power consumption of the terminal device more accurately,this thesis has developed a test device for analyzing the power consumption of a terminal device.The terminal device decides the computation offloading of the deep learning network according to RSSI,the terminal device and the edge server respectively run the client and server programs to complete deep learning network layer calculation,the power consumption testing device also records the terminal device power consumption.At the end of the experiment,the statistical analysis of the differences in the power consumption of the terminal device under three different modes of RSSI computation offloading,pure local computation and pure edge server computation under different RSSI was performed.Experiments show that when the RSSI of this terminal device is around-50 d Bm,the network quality and the device throughput are relatively good,offload all computing tasks to the edge server can reduce power consumption by 35%,when the RSSI is around-70 d Bm and the terminal device throughput is reduced,using the offloading module to offload part of the deep learning network to the edge server can reduce the power consumption by 22%,when the RSSI is-80 d Bm,the device throughput is very low and the communication overhead between the terminal device and the edge server increases.The use of computation offloading will reduce the power consumption of the terminal device within a specific RSSI range.When the RSSI strength is low,the power consumption of local computing is lower than that of edge computing.In this case,computation offloading will increase terminal device power consumption.